Technical note: an R package for fitting sparse neural networks with application in animal breeding

Area of Science:

• Animal Breeding and Genetics
• Computational Biology
• Machine Learning in Genomics

Background:

• Neural networks (NNs) show promise for genomic selection (GS) in animal breeding.
• Over-parameterization and high-dimensional input hinder NN application in GS for predicting phenotypic outcomes.

Purpose of the Study:

• To develop an R package, snnR, for optimizing NN structure in GS.
• To address over-parameterization and improve prediction accuracy using sparse NNs.

Main Methods:

• Developed the snnR R package implementing sparse NNs.
• Minimized square error with L1-norm penalty for optimal NN structure.
• Compared snnR with the brnn R package using genotype or genomic relationship matrices.

Main Results:

• The snnR package effectively finds optimal sparse NN structures.
• snnR significantly improved computational efficiency compared to brnn.
• snnR demonstrated enhanced prediction ability for GS in animal breeding.

Conclusions:

• The snnR package offers a feasible and effective solution for GS.
• Sparse NNs with multiple hidden layers improve computational efficiency and prediction accuracy.
• snnR advances the application of NNs in animal breeding for genomic selection.